Agentic framework for intent-driven responses in computer-based mortgage systems

This invention relates generally to computer based systems. More specifically, at least one embodiment, relates to an agentic framework for intent-driven responses in computer-based mortgage systems.

Today, mortgage industry operations rely heavily on computer-based systems, in part, because of the huge amount of information that must be sorted and processed in a search. The problem has increased over time due to the variations found among mortgage offerings, the number of competing products and the ability for these offerings to slightly differ based on the borrower's qualifications and the property. These computer-based systems allow loan officers to evaluate possible mortgage product options for any given request where it is impractical to perform the analysis in the human mind. However, current approaches suffer from problems that limit their effectiveness. One common approach is to provide these systems with rigid rules-based decision trees that are used in searches for a desired product. This approach lacks the flexibility to properly process queries given the size of the information set and dynamic nature and nuance found in queries. Other current tools are expected to operate in a highly autonomous manner where they are solely responsible for a wide range of complex operations. In practice, these tools provide results that can be unpredictable. For example, large language models (LLMs) are being introduced into mortgage systems. However, LLMs are generally stochastic in nature. That is, there is an element of randomness to their responses even when they receive the same natural language input.

In addition, mortgage product and rate data provides a vast, multi-dimensional data set that traditional systems cannot effectively search when presented with a natural language request. These natural language queries are often nuanced because they provide a set of objectives presented in a form that includes objectives that may compete with one another. For example, a user may offer “I need a $5K rebate combined with the lowest rate and a price close to par.” These queries can include a mix of quantitative and qualitative criteria that require logical leaps. Traditional systems struggle with these types of mortgage product and rate data searches queries due to the vast amount of information present in today's mortgage rate stacks and because these queries include requirements that are not well suited to rigid filters or algorithms that seek exact matches.

In addition, traditional AI-based approaches can overwhelm conventional reasoning models with complete sets of data, for example, mortgage product records that can number in the hundreds of thousands. That is, presenting a large language model with so much information that it is unable to hold attention/context when performing such a complex task.

Embodiments described herein overcome the above-described problems found with conventional computer-based mortgage systems. For example, various embodiments provide an agent-based AI system including intent classification. An intent classifier includes a large language model that operates to classify an intent of a user query and identify an action handler to orchestrate operation of the system using a predefined workflow. The multi-agent system includes a plurality of agents each including an associated large language model. The individual large language models can be individually fine-tuned with examples to optimize the respective agent for the reasoning task(s) to which it is dedicated. The action handlers designated for specified actions provide the logic for the defined workflow to complete the respective action using the multi-agent system. The structure of the overall agentic framework provides a dynamic toolset that allows a query to be resolved through a series of actions performed by specialized agents assigned to a task by the action handler. This structure provides more consistent operation by reducing the complexity of the individual task(s) or sub-task(s) performed by each agent. The system also provides a discrete focus that provides more consistent results because the action handler has a predefined set of tasks for completion by a predefined set of agents in a predefined order to resolve a user's query. This resolves queries using large language models to deliver deterministic results.

The agent-based AI system also includes a generalist agent that is employed to respond to user messages for which a specific intent is not identified by the intent classifier. This provides the system with an ability to respond in context to messages for which only a general intent is identified by the intent classifier. The use of a conversational agent or chat widget provides the system with an ability for an on-going dialog with the user that can result in an identification of a specific intent, in a form of “self-healing” operation.

Embodiments of the AI agent-based system include a semantic matching agent to resolve user queries with precise, context aware and user-friendly results for mortgage product and mortgage rate data searches. These embodiments can employ a set of AI agents each including a respective large language model that is fine tuned for the reasoning task that it is assigned in a workflow that identifies the proper data from within a mortgage rate stack that reflects what the user is asking for. Some of these embodiments operate to search data that is organized in a tabular format, for example, spreadsheets and .csv files, (that is, a known file-type) to provide refined highly responsive replies to mortgage product and mortgage rate data queries that are impossible to parse for traditional computer-based mortgage systems. The operation of the AI agent-based system is flexible and unconstrained by rigid matching algorithms and can provide results that may not provide an exact match but are relevant in context because they include solutions that are close to meeting the conditions established by the user's query.

Unlike conventional AI approaches, embodiments described herein employ one or more AI agents in operation of a tournament-based semantic match on mortgage records that are included in a tabular mortgage “rate stack.” These embodiments systematically organize qualifying mortgage records into subgroups that are concurrently evaluated by the large language model. This process operates to continuously reduce the total quantity of qualifying records and optimize performance by selecting a subset of records from each subgroup at each round based on those records that most closely match the user's mortgage product and mortgage rate data queries. This dynamically reduces the search space while maintaining the context of the query. Results are provided with a large language model that explains the best matching mortgage records using natural language and in the context of the original query.

As used in herein, the terms “semantic match” and “semantic matching” are used to describe a reasoning-based search that leverages the capabilities of a large language model to identify a similarity between information available in tabular data to elements of a user's query while considering inter-related factors that impact the suitability of a potential match. One of ordinary skill in the art in view of the disclosure herein will recognize that a search of quantified data included in a set of tabular data with a reasoning-based approach, considering the factors, to identify and rank records that best match the requirements and objectives (that is, what the user is asking for while considering these inter-related factors) included in a text-based user query provides a “semantic match” or “semantic matching” as the terms are used herein. One of ordinary skill in the art in view of the disclosure herein will also recognize that an approach that searches a vector database utilizing some form of vector based comparison or search method, for example, a nearest-neighbor algorithm, lacks this reasoning and therefore does not perform “semantic matching” or provide a “semantic match” as the terms are used herein.

This invention is not limited in its application to the details of construction and the arrangement of components set forth in the following description or illustrated in the drawings. The invention is capable of other embodiments and of being practiced or of being carried out in various ways. Also, the phraseology and terminology used herein is for the purpose of description and should not be regarded as limiting. The use of “including,” “comprising,” or “having,” “containing”, “involving”, and variations thereof herein, is meant to encompass the items listed thereafter and equivalents thereof as well as additional items.

Embodiments of the agentic-based AI system described herein provide a dynamic framework with which user queries can be effectively resolved. The intent classifier as described herein allows for these queries to be processed with the system having a clear understanding of the problem to be solved based on the user's intent as conveyed in the query. An identification of the assigned action handler at the front end of the process results in operations performed based on an early identification of tasks and sub-tasks that should be performed to effectively respond to the query. This also limits autonomy that is allowed the large language models employed in the system that might otherwise establish a less effective set of operations to resolve the query. The action handler responsible for resolving the query operates to provide API calls to additional agents and other function calls as needed to resolve the query and return a natural language response. In this role, the action handler which can be identified based on an identified intent, orchestrates operation by other agents to complete sub-tasks, receive resulting data and other information returned to the action handler from the agents and other system resources, and provide follow-on function calls as needed to resolve the user's query and generate a natural-language response.

The technical advantages provided by the embodiments described herein are achieved by employing the action handlers to identify specific actions and sets of actions to complete within the overall framework of the agentic-based AI system. This approach provides the benefits of structured planning to coordinate overall operation while allowing the reasoning capabilities of the large language models to be fully leveraged in a more focused way. That is, by planning activities for specialized agents in advance for a given set of tasks, the agentic framework overcomes technical problems found with systems that give their LLMs full autonomy including freedom to define how to complete tasks in real time. In summary, the embodiments described herein leverage the capabilities of large language models in a manner that delivers deterministic results not typically produced by other more autonomous LLM frameworks.

The identification of these actions also reduces the complexity of the various tasks and sub-tasks required to be performed by any one of the large language models included in the agentic-based AI system. Less costly large language models can be employed because the complexity and the breadth of the tasks and sub-tasks they perform are reduced. This reduces the computational load placed on each agent. In addition, the on-going availability of the chat widget in combination with the intent classifier provides the system with an ability to maintain a conversational dialog in context. This allows the system to adapt to the substance of the user query and identify a user's intent or provide an alternate relevant natural-language response even if a specific intent cannot be classified.

In various embodiments, operation of the AI agent-based system is flexible and unconstrained by rigid matching algorithms and can provide results that may not provide an exact match but are relevant in context because they are a close solution. For example, a user message may establish that “I need a payment no more than $1500 month.” Traditional sorting techniques system apply “hard” filters with fixed requirements that lack context. Faced with the preceding user message, a traditional approach will eliminate a result that is even just a dollar outside the requirement. In this case, an eligible mortgage product that requires a payment of $1501 will be excluded from the results. In contrast, embodiments, described herein perform sorting with a large language model that is fine tuned to use discretion to find useful results. Therefore, given the user message provided above, embodiments described herein will include a result with a payment of $1501.

The intelligent sorting algorithm described herein solves the technical problems found with traditional rules-based sorting algorithms because it employs a semantic match sorting that includes qualitative criteria rather than just being based on integer values. This improves the search results because it captures the nuances of the user's intent to deliver the most relevant options. Semantic sorting is supported through operation of multiple different agents with separate large language models trained to tag and classify the features of the mortgages of interest based on the user query, and to identify the locations of the information within the records included in the data set, respectively. Some embodiments include tournament-style semantic match sorting that evaluates subsets of eligible mortgage products and combines the best matching results in each subgroup in an iterative process that reduces the quantity of candidate-records at each round. The process continues until a threshold quantity of records is reached. A final semantic matching round is completed to select the best matching from among this final subgroup of “winners.” Including qualitative search criteria provides the large language model that performs the semantic match process with discretion to find all useful results even those that are not precise matches. The overall agentic framework and workflow also provides an advantage because a conversational response can explain nuanced results in context, for example, what results are being shared and why.

Referring to, a flow diagram of a processperformed by an agent-based AI system is illustrated in accordance with various embodiments. The processprovides a general overview of operations performed by the system to illustrate at a high level the role of system elements including the intent classifier, the complete set of agent-based tools and the AI-based chat widget. A portion of an example session is provided in Table 1 to provide context for the description of the process.

The example session illustrated in Table 1 includes separate rows for: communication received from the user by the chat widget; output messages provided by an agent-based AI system; and responses generated by the chat widget following execution of a set of tasks performed by the agent-based AI system. For clarity, only some of the activities of the agent-based AI system are included in Table 1. Some of the activities that are not included in Table 1 are described further below with reference to the process illustrated in. The session illustrates the use of natural language processing employed in response to user queries.

The processbegins with a receipt of a message input received from a user at block. The message(s) received from the user are input by the user from a user computing device in a natural language. The message input is provided to a chat widget included in the front end of the agent-based AI system which operates to generate a text input for the large language model included in the intent classifier. The intent classifier operates on the user's message as provided in the text input to perform an intent classification operation at block. The intent classifier operates to identify an intent and an action corresponding to the intent, for example, explain ineligibility, run a near miss scenario, update a form. The intent classifier returns an action key that corresponds to the action handler designated to handle the action and initiate a predetermined set of tasks to be completed to respond to the user's message at block. The designated action handler is configured to leverage the agent-based AI system and other resources by orchestrating a set of operations that can include agent operations utilizing the respective large language models at block. In various embodiments, a single agent or a set of agents included in the agent-based AI system operate to perform the tasks and sub-tasks required to respond to the query at block. The operations completed at blockcan also include other function calls at blockthat call other non-agent functions and other resources to complete tasks needed to provide a response message. The response message is provided in a natural language and presented to the user via the chat widget at block. According to various embodiments, response messages can take two general forms depending on the results of the intent classification at block.

For example, Table 1 illustrates an intent classification that finds a specific intent. Here, following an introduction from the chat widget the user provides a message “we want the lowest rate but the borrower can only afford to buy down up to $5K.” At block, the intent classification large language model outputs a system message “explain pricing.” The system message identifies a specific intent to the backend that also identifies the respective action handler that is designated to manage operations that provide the output message. In this example, the action handler communicates information from the user message to a normalization tagger agent (block). The normalization tagger agent is the first of multiple agents that are employed together to resolve this user query and return a reply in context. Here, the normalization tag or agent outputs tags to place the user's query in a standard format and tag it with descriptors. The descriptors can include requirements that must be met and improvements or “optimizations” that are desired but not required. Here, at Step, the normalization tagger replies with the tags “lowest RATE” and “CREDIT/COST around 5000.” Each of these include the descriptor (OPTIMIZATION) to indicate that these criteria are desired by the user but are not required. That is, the optimizations are objectives or goals but are not required.

These and other operations are completed at blockof, before a response message is generated at block. Because the intent classification identifies a specific intent at block, the response is based on the specific intent at block. The example in Table 1 illustrates a response message that provides a summary of rates with the original context “Here's a summary of the rates tailored to your criteria of seeking a low rate with the ability to pay around $5,000 for buying down” that is delivered to the user via the chat widget.

The system includes a self-healing aspect that also allows it to provide response messages even where the intent classification at blockonly identifies a general intent. For example, if a user provides a query that lacks sufficient information to identify a specific intent, the system can provide a response message at blockthat either provides the best available general response or respond with a query of its own. The capabilities of the chat widget in combination with the agent-based AI system deliver a natural language dialogue to the user. According to some embodiments, the agent-based AI system includes a domain generalist agent, for example, a mortgage assistant. The mortgage assistant can be employed to answer general questions on mortgages and mortgage pricing where specific intent is not found at block. One example of these capabilities is in circumstances where the user message provides a question that lacks specifics. For example, a user faced with a set of results for a family of mortgage products such as a 30 year fixed mortgages may inquire “why is the borrower ineligible for the 30 year fixed products?” Here, the mortgage assistant agent is employed at blockto generate a Q&A agent response “I can assist. Please let me know the name of the product you are interested in.” The response received from the user in reply to the follow-up is treated as a new message at block. With sufficient information, such as an identification of the product, a specific intent can be identified at blockrouted to the appropriate action handler at blockto coordinate back-end operations at block. The result of these actions is the delivery of a response based on specific intent at block.

Referring now to, a system level block diagram of an agentic frameworkof an agent-based AI system is illustrated in accordance with various embodiments. The agentic frameworkincludes an agentic frontendand a multi-agent systemthat includes a plurality of AI agents each dedicated to performing a respective task or set of tasks within the overall operation of the agent-based AI system. The intent classification included as a part of the multi-agent systemallows the systemto provide an action handler that is dedicated to completing the overall task with a set of operations that are known from the start. This approach provides the benefits of structured planning to coordinate overall operation for a known intent while allowing the reasoning capabilities of the large language models to be fully leveraged in a more focused way. That is, by planning activities for specialized agents in advance for a given set of tasks based on intent, the agentic frameworkovercomes technical problems found with systems that give their LLMs full autonomy including freedom to define how to complete tasks in real time. In summary, the embodiments described herein leverage the capabilities of large language models in a manner that delivers deterministic results not typically produced by other more autonomous LLM frameworks.

According to the illustrated embodiment, the frontendincludes an action router, action handlers, a response handler, and a chat widget. The chat widget is employed to receive user message inputs and return message outputs from the multi-agent systemand to do so in a conversational manner in a natural language. Although different categories of users having various roles can employ the agentic frameworkdepending on the embodiment, a user having the role of a loan officer is described in the examples provided herein. The front endorchestrates operation of multi-agent systemwith an action handlerselected to handle the user's message communicating with the other agents within the multi-agent system. In general, the communication consists of function calls from the action handlerand API responses from the agents. The action handlersare also configured to communicate with additional resourcesthat are accessed by the agentic framework to perform the tasks required to complete the action identified by the intent. For example, function calls can be made to backend services included in a web hosted mortgage system as illustrated and described with reference to. The additional resourcescan also include third party resources such as generalist large language models and large language model fine-tuning systems hosted by third parties. The function calls can include API calls made to the backend services or third party resources and retrieval of API responses for processing with the front endand/or elements included in the multi-agent system.

The front end also includes an action routerand a response handler. Where an action corresponds to an identified intent, the action routeroperates to activate one of the action handlersbased on the intent received from an intent classifier included in the multi-agent system. The response handler operates to generate a conversational output for the chat widget from information received by the action handler. For example, an output from a large language model received by the action handler in an API response returned from the multi-agent system.

In the illustrated embodiment, the multi-agent systemprovides a decentralized agent network that includes agents dedicated to specialized functions within the system. Each of the agents includes a large language model that is trained in designated tasks and operates to deliver an output required to complete the task. In embodiments, some of these agents can include a foundational large language model with a system message unique to that agent. The system message provides the base operating instructions for the agent including an identification of the task it is to perform, the type of input it can expect to receive, and what the agent is expected to do with the input (for example, as established in instructions or a set of rules). Some of these agents can also include large language models that are trained in a fine-tuning process. The fine tuning can include training data including sets of: system messages; user messages; and output messages for that agent. The training can include the use of session data or synthetic data that is specifically created to train the model with realistic examples.

As described in more detail below, an intent classifier agentenables the decentralized operation with an identification of a responsible action handler based on the content of the user's message. Among the multi-agent system, a mortgage assistant agentprovides an agent that is trained for a variety of more generalized tasks for a specific domain, here mortgage operations. The multi-agent systemincludes an option selector agentto select an option from a set of options. Specifically, the option selector agentis trained to select an option from among a set of possible options each associated with a unique ID. The option selector agentreturns the unique ID for the selected option.

A voice generator agentand an audio transcriber agentare included to facilitate dialog in a natural language. For example, the voice generator agentis a text-to-speech machine learning model that generates an audio stream based on the given text. This is used to enable the chat widget to provide an audio response in a natural language. The audio transcriber agentis a speech-to-text machine learning model that generates a text transcription of an audio file. This is used to allow the user to dictate messages to the chat widget. A column selector agentis provided for use with certain file types that are sources of information used to resolve user queries. A contextualizer agentoperates in combination with the chat widget for example where retrieval-augmented generation (RAG) is employed to generate natural language responses. The systemalso includes a JSON translator agentthat operates to output JSON based on the text included in the user's message, a name extractor agentwhich operates to extract the name or subject from a message, and a summarizer agentto summarize the text of messages processed by the system, for example, with longer messages before storing conversation history to reduce overall token usage. A set of agents employed to generate explanations to user messages are also included. In the illustrated embodiment, these include a near miss explainer agent, a rate explainer agent, and an ineligibility explainer agent.

In operation, the intent classifier agentreceives the user message received from the chat widget and determines an intention and, if an action matches the intent, provides an action key that is employed by the action routerto route the action to the appropriate action handler to handle the request. Each of the action handlersdefines a particular workflow including a predetermined set of tasks. The workflow also includes an identification of any agents included in the multi-agent systemthat are utilized to complete the set of tasks, and the order in which those agents are employed in the workflow.

Each of the action handlerscan be dedicated to a selected action in reply to known categories of user queries including, for example, product inquiries, rate inquiries, and eligibility inquiries. The action handlersoperate with a known workflow that leverages the multi agent systemthrough function calls including API calls or calls for other functions available to agentic frameworkeither within the framework or external to the framework. In general, action handlersoperate as intermediaries between the various agents to share information. The workflow can include one or a plurality of agents depending on the action. The use of action handlers dedicated to one selected action combined with the operation of the multi-agent systemprovides a dynamic and highly responsive system. Depending on the action, the action handlersmay require operation of one or a plurality of the agents included in the multi-agent system. Where multiple agents are employed to complete a given action they may operate sequentially or in parallel. For example, the action handlerscan operate to receive an API response from a first large language model included in a first agent and pass that information as an input to a second large language model included in a second agent. When the action (a task or set of tasks) is completed, the action handlerspass the information to the response handler. The response handleroperates to generate a conversational natural language output that describes the outcome of the action. This can include associated data that is returned, for example, data concerning mortgage products, mortgage rates and mortgage eligibility that are included in a conversational reply. Depending on the embodiment, the quantity and type of agents included in the multi-agent systemcan differ.

Additional details concerning some of the agents included in the multi-agent systemare provided here. For example, the mortgage assistant agentprovides a tool that is used to reply to user messages with answers to mortgage questions, explaining pricing, answering questions about the assistant's capabilities, and as the default conversation bot when no other actions are specified. This provides the agentic frameworkwith a domain generalist that can also redirect conversations to other agents in the multi-agent systemwhen, because of a conversation with the user, it identifies a specific action that will assist in providing a more complete reply. Trained for this role, the mortgage assistant can be employed to answer general questions concerning mortgages where a specific intent is not found by the intent classifier. In various embodiments, the mortgage assistant agentcan be molded to perform a variety of tasks via post parameters such as, “Additional Rules” that allow a developer to add additional behavior rules to the agent's system message prompt, or “Additional Context” which allows the developer to include task specific reference material for the tool.

The column selector agentis used to select columns from electronic files, such as .csv files that organize data in a tabular form. As is described in greater detail below, the column selector agentcan be employed to select the columns with data that is most relevant to generate a reply to a user's question. This is advantageous because it can be used to filter for higher value content and reduce the complexity and volume of data that must later be analyzed by or explained to other LLM-based tools including the other agents. According to other embodiments, other dedicated agents can be included for use with file types and databases that organize information differently, and in non-tabular formats. For example, the multi-agent systemcan include a key selection agent including a large language model to operate with JSON, YAML and other human readable data serialization formats to select (or filter) out specific keys from the JSON or YAML object in response to a text query.

The contextualizer agentuses conversation history to find and replace things like pronouns or abbreviations with the actual information being referenced. This makes the user's query more semantically rich and accurate. This can assist when querying against things like vector databases, where the query should have as much semantic information embedded in it as possible.

The normalization tagger agentis used to normalize the user's query to a standard format, and as illustrated in Table 1, is employed in tasks such as a rate stack search where it is used to categorize the criteria provided by the user. For example, the normalization taggercan categorize a user's criteria into those criteria that are required to meet the user's objectives provided in the query and those criteria that, if available, will optimize the result for the user. In some context, the required criteria are referred to as “hard requirements” and the criteria that will improve the results for the user are referred to as “optimization”.

The JSON translator agenttakes a user message and a translation schema, then extracts values and returns minified JSON based on the provided translation schema. The JSON that is output can be used in a variety of ways including updating forms on the frontend, or making API requests.

The name extractor agentextracts the name or subject from a message. The name extractor agentis used where operations include saving or loading a loan scenario if the name of the scenario needs to be extracted from the user message and supplied to the relevant API endpoint or function.

The semantic match agentoperates to find the best matches in response to a user's natural language query. As is described further below with reference to, the semantic match agentis advantageously employed in combination with the column selector agentand the normalization tagger agent. In some embodiments, the semantic match agentoperates on structured data, like .csv files.

The near-miss explainer agentis employed with scenario analysis performed by agentic frameworkto identify mortgage products for which a borrower almost qualifies. The near-miss explainer agentoperates to provide information that explains to a user the product or products that they may qualify for and what they must do to qualify.

The rate explainer agentoperates to provide information that identifies mortgage products with associated details such as the type of product (fixed rate versus adjustable rate), the interest rate of the product, the term of the product and associated costs as some examples. In some embodiments, the rate explainer agentis employed in combination with the results of a rate stack search to provide the response handler with information that can be included in a natural language response message via the chat widget.

The ineligibility explainer agentis employed with scenario analysis performed by the agentic frameworkto identify mortgage products for which a borrower is ineligible. In these embodiments, the ineligibility explainer agentcan operate to provide the single most important (that is, the easiest to address and/or the most impactful) reason why a mortgage product is ineligible for the borrower. The ineligibility explainer agentcan also operate to provide the most important reasons among multiple reasons why a mortgage product is ineligible for the borrower.

Referring now to, a network operating environment for implementing a systemincluding an agentic framework as described herein is illustrated in accordance with various embodiments. The systemincludes a mortgage systemas an example to describe the functionality and operation of the system. However, those of ordinary skill in the art will recognize based on the disclosure provided herein that the systemcan be employed in a similar fashion in other fields. The systemalso includes an end user devicethat is a representative example of an end user device where each of a plurality of end users (for example, loan officers) employs their end user device, respectively, to access the resources provided by the mortgage system. The systemcan support access by any number of end user devices. Depending on the embodiment, the systemcan also integrate external resources, for example, large language models(including third party general purpose LLMs) and third-party LLM fine tuning systems.

According to the illustrated embodiment, the mortgage systemincludes an agent-based tool setthat includes an intent classifierand a plurality of AI agents. The quantity of AI agentscan vary and can include any number of AI agents depending on the embodiment (represented here as Agents-N). The intent classifierand the plurality of AI agentseach includes a large language model, respectively. The mortgage system also includes a processor, a network interface, I/O, data storage, and a memory. In general, the memorystores computer programsthat include software instructions for execution by the processor. The processorcan include one or a plurality of processors.

Depending on the embodiment, the mortgage systemcan include one or more of a variety of computing devices such as a general purpose computer such as a PC, a laptop, a tablet computer, or other computing device. The resources included in the mortgage systemcan be hosted on one or more servers accessible to the system operator and users via the network. In various embodiments, the resources included in the mortgage systemare hosted on servers located at one or a plurality of sites. The mortgage systemcan be hosted by the entity that delivers the services and resources provided by the mortgage system or by a third-party server host.

The network interfaceis employed for communication between the mortgage systemand other elements connected to the networkincluding the end user computing devicesand the external resources,. In general, the networkcan include either or both of local-area networks (LANs), wide area networks (WANs), wireless communication, wired communication and may include the Internet. The networkprovides access to one or more remote devices, servers, application resource management and/or data storage systems. For example, the networkcan allow communication between any of a plurality of end user devicesand the mortgage system. In general, the systemprovides for communication of the illustrated components with one another and/or with any of the other resources and devices coupled to the network. Communication can occur using any of Wi-Fi networks, Bluetooth communication, cellular networks, satellite communication, and peer-to-peer networks available either alone or in combination with one another via the network. Depending on the embodiment, the networkmay be any type and/or form of network known to those of ordinary skill in the art capable of supporting the operations described herein. Thus, other communication protocols and topologies can also be implemented in accordance with various embodiments.

The memoryprovides a non-transitory machine readable storage media which is coupled to the processor. The memorycan store the computer programsthat when executed by the processorprovide for operation of the agent-based tool setand delivery of a front end to users on their end user device. The front end allows for an interactive set of operations by which the user can access and utilize the agent-based tool set, for example, via a conversational dialog available through a display included in the end user device. According to the illustrated embodiments, the programsdirect all aspects of the agentic system operation.

The I/Ocan include any of the display in which a user interface is presented to the user (for example a developer or a system administrator), a keyboard, a mouse, a touchscreen controller where, for example, the display is a touchscreen display, or alternatively, a trackpad or mouse used to move a cursor within a user interface such as a GUI. According to further embodiments, the I/O can include an audio system employed with a speech recognition system to allow hands-free interaction with the GUI.

According to some embodiments, the data storagestores information concerning various aspects of the mortgage system, for example, information on mortgage products, mortgage rates and mortgage costs. Depending on the embodiment, the data storagecan include any of a relational database, object-oriented database, unstructured database, or other database. Further, the data storagecan be included in any aspect of a memory system, such as in RAM, ROM, or disc, and may also be separately stored on one or more dedicated data servers included in the mortgage system. The elements included in the mortgage systemare coupled to one another with an internal communication system (not illustrated) to allow for a transfer of data, execution of commands, or exchange of messages within the system.

According to the illustrated embodiment, the end user deviceis a computing device that includes a user interface, for example, a graphical user interface. In overall operation, the systemis employed to provide a set of services to the respective end users as is described herein. The end user devicecan include any type of computing device suitable for communicating with the mortgage system via the network. Accordingly, the end user devicecan include one or more of a variety of computing devices, for example, a general purpose computer including a mobile phone, a PC, a laptop, a tablet computer or other computing device. The end user deviceis representative of the devices employed by end users to access the resources provided by the mortgage systemas well as any other external resources employed in the system.

The end user deviceincludes an agentic system front end, a user interface, a processor, a memoryand a communication interface. According to the illustrated embodiment, the agentic system front endincludes a chat widgetand action handlers for agent integration. In general, the action handlersare utilized to coordinate the execution of tasks performed by the agent based tool set, for example, by generating API calls to agents and handling API responses as needed to complete a workflow.

The elements of the end user devicecan be configured like corresponding elements of the mortgage system. Depending on the embodiment, the user interfacecan include a display in combination with a keypad, a mouse, a touchscreen, a speaker, and/or a microphone to allow the user to interact with the agentic system front end. The communication interfacecan include a wired or a wireless interface, for example, including a Wi-Fi and/or BLUETOOTH system to enable communication with other elements of the systemover the network. The processorcan include one or more central processing units (CPUs), graphic processing units (GPUs), field-programmable gate arrays (FPGAs) or application-specific integrated circuits (ASICs) either alone or in combination with one another depending on the embodiment. The memoryincludes a non-transitory machine-readable storage media which is coupled to the processorto provide information, data and instructions accessible for use and, in the case of instructions, execution by the processor.